<h1 id="v1.0">v1.0</h1>CyberRealistic Pony merged with Big Lust, and a dusting of GonzaLomo.This 'redux' version has no additional fine-tuning, and the full merge recipe workflow is included in the metadata.<h3 id="notes">Notes</h3>For character and concept LoRA training, I have had excellent results using Big Lust v1.6 as the base model with settings:<ul><li>resolution 1024</li><li>rank/alpha 128</li><li>LR 2.5e-5</li><li>cosine scheduler w/250 steps warmup</li><li>optimi lion optimizer</li><li>batch size 8</li><li>and between 100 - 1000 LLM tagged images w/masks. I use Qwen3-VL-30B-A3B-Thinking</li><li>~step 1250 is the sweet spot</li></ul>My showcase images are not actually using the DMD2 LoRA, however it is present in the workflow in an inactive branch which is confusing the Save Image w/ Metadata Universal node. The model does work with DMD2 but I am not a fan. I recommend using strength 0.7, CFG 1.3, 14 steps, LCM sampler with it if that's your thing.I have not yet fully tested this for compatibility with Pony &amp; SDXL LoRAs.TBH I don't like the structure or fine details produced by this model as compared with my banned versions. Some work is required to fine-tune it. I am working on undoing some of the NSFW-by-default bias coming from Big Lust by fine-tuning this model prior to merge.Adding GonzaLomo was likely a mistake; my impression is it seems to be overbaked with compression artifacts and is heavy handed with adding watermarks to gens. The merged model then seems to fuck up hands more than my own fine-tunes. TL;DR filter and pre-process your datasets. I have tested candidates for the next version with less/no GonzaLomo and I like the results more.

Files with SHA256: b35fbaa074e093fbad51a65855cccbf63f6b80fd257d9492fc538b51090c77ae